摘要
工业产品在现代社会中无处不在,但在工业生产过程中不可避免地会出现一些产品质量问题,针对工业场景下收集的样本图像类间数据量不均衡、总体样本量较小等问题,引入上下文关系编码器(CRE)与孪生神经网络(SNN)结合的CRE-SNN工业缺陷分类网络。该网络不需要对样本数据进行标注,经过循环迭代生成标注的特征图像,能够解决标注耗时的问题;同时,该网络使用对特征对比的方式进行分类,可以应对工业场景下样本数量不足的问题。训练出的网络能够应用于多种不同的数据集和仅有几张样本图像的极端情况,具有良好的泛化性和鲁棒性。在3个不同工业数据集上的实验表明,与其他经典算法相比,所损网络的准确率、平均准确率(mAP)等指标得到了一定的提升。
Industrial products are ubiquitous in modern society,but some product quality problems inevitably occur in the industrial production process.To address the problems of unbalanced data volume between classes of sample images collected in industrial scenarios and small overall sample size,this paper introduces a CRE-SNN industrial defect classification network combining contextual relation encoder(CRE)and siamese neural network(SNN),which does not require labeling of sample data and generates labeled feature images through cyclic iteration.The model can solve the problem of time-consuming labeling.Meanwhile,the network adopts a feature contrast-based classification approach,which can address the challenge of insufficient sample quantities in industrial scenarios.The trained network can be applied to many different datasets and extreme cases with only a few sample images,and has good generalization and robustness.Experiments on three different industrial datasets show that the accuracy and mAP of the proposed network are improved compared with other classical algorithms.
作者
张卉婧
李敏波
ZHANG Huijing;LI Minbo(School of Software,Fudan University,Shanghai 200433,China)
出处
《微型电脑应用》
2025年第4期287-290,共4页
Microcomputer Applications
关键词
缺陷检测
上下文关系编码器
连体神经网络
小样本学习
defect detection
contextual relation encoder
siamese neural network
few-shot learning
作者简介
张卉婧(1998-),女,硕士研究生,研究方向为大数据与数据科学。;通信作者:李敏波(1970-),男,博士,副教授,研究方向为大数据与数据科学。
